Inside the Kernel | Exploring GBase 8c: The Database Kernel Is More Fascinating Than You Think
The China Academy of Information and Communications Technology (CAICT), in its Database Development Research Report (2026), concludes that database technology is shifting from "cloud-native" to "AI-native" — with AI capabilities evolving from bolt-on tools into native capabilities embedded in the kernel. In response to this trend, this column, "Inside the Kernel — Exploring the GBase 8c Kernel World," systematically breaks down GBase 8c's kernel design, explaining the implementation principles and real-world value of each capability. This first installment outlines GBase 8c's kernel positioning and overall architecture.
Why Talk About the Kernel?
CAICT's report projects China's database market will approach RMB 98 billion by 2028, with a CAGR of approximately 13%. According to CAICT research, about 20% of leading enterprises have already deployed database agents in development or operations scenarios, while more than half remain in the planning and pilot stage. "Database + AI" has moved from proof of concept into core system selection.
Yet during the selection process, customers tend to focus on questions like "does it support vectors," "is it Oracle-compatible," or "how does it score on benchmarks." These metrics matter — but once core business actually migrates in, more fundamental questions surface: Why do mixed workloads contend for resources? Can slow SQL be solved by adding indexes alone? During online scaling, how should data be redistributed to avoid jitter? The answers to all of these live in kernel design.
There is plenty of content on the market comparing products, but little on kernel design trade-offs. This column exists to fill that gap.
"AI-Native" Is Not a Bolted-On AI Assistant
According to CAICT's definition, AI-native means using AI capabilities as the foundation throughout product design, architecture, and business processes — rather than layering AI plugins onto traditional systems.
Applied to GBase 8c's design, this means the query optimizer must sense workload types and automatically select execution paths; the storage engine must natively accommodate mixed vector and scalar data processing; the operations framework must provide fault self-detection, self-diagnosis, and self-healing; and the security module must enforce access control based on data semantics — not merely offer an "intelligent Q&A" entry point in the UI.
Industry research frames this evolution in five stages: AI-suggested, AI-assisted, AI-enhanced, AI self-assembling, and AI self-designing. The industry as a whole is currently transitioning from "assisted" to "enhanced." GBase 8c's goal is anchored here: through three core capabilities — intelligent optimization, intelligent operations, and intelligent security — to drive databases from "manual tuning" toward "self-tuning and self-healing."
The GBase 8c Kernel at a Glance
GBase 8c's kernel design can be summarized as "one skeleton, two foundations, three layers of intelligence."
The Skeleton: Native Distributed Architecture
GBase 8c does not assemble a distributed middleware layer atop a single-node database. Instead, it builds distributed capabilities natively — from storage and transactions to query optimization. Cross-node transactions use two-phase commit (2PC) with supporting optimizations, and rely on globally consistent timestamps from the Global Transaction Manager (GTM) to guarantee strong consistency: every distributed transaction either commits entirely or rolls back entirely, with no intermediate state where some nodes commit and others fail. Data distribution supports Hash, Range, List, and other sharding strategies, combined with replicated and distributed tables, so that join queries execute efficiently across nodes.
Foundation 1: Multi-Model Storage
A single database natively supports multiple engines:
Row-store engine for high-concurrency OLTP transactions
Column-store engine for wide-table scans and complex aggregation
Row-column hybrid storage, giving a single table both physical forms, with the optimizer automatically choosing the optimal path — transactional writes go to row-store, reporting analytics go to column-store, with no ETL movement in between
It also includes a built-in time-series engine (automatic partitioning, continuous aggregation), a GIS engine (native distributed spatial database capabilities, running stably in national-level systems for years), and a vector engine (supporting up to 64,000-dimensional vectors, with mainstream index algorithms including HNSW, and hybrid vector-scalar retrieval within a single SQL statement).
Foundation 2: Multi-Form Deployment
The same kernel adapts to business scale: from minimal single-machine deployment, up to primary-with-multiple-standbys, to fully redundant distributed deployment. The most elastic form is storage-compute separation — a stateless compute layer mounted on shared object storage, supporting second-level scaling to handle business peaks. This architecture also addresses the access patterns of the AI agent era: agent concurrency differs from traditional business traffic, and only storage-compute separation can effectively support bursty, irregular access peaks.
Three Layers of Intelligence Across Both Foundations
Intelligent optimization delivers adaptive query plans and slow SQL remediation
Intelligent operations covers fault prediction, diagnosis, and full-lifecycle monitoring
Intelligent security achieves EAL4+ certification, providing full encryption-in-use, transparent encryption, tamper resistance, and dynamic data masking
In addition, GBase 8c offers several key capabilities: a high-availability and disaster recovery framework supporting same-data-center, same-city dual-center, two-city three-center, two-city four-center, and cross-region active-active architectures, with RPO = 0 and sub-second RTO. On the ecosystem side, it is highly compatible with PostgreSQL syntax and highly compatible with Oracle, MySQL, and SQL Server, with automated migration tools that minimize the cost of adapting existing business systems.
Based on this framework, we have identified eight core focus areas for GBase 8c: multi-model multi-form storage, hybrid vector-scalar retrieval, intelligent optimization, intelligent operations, distributed consistency and high availability, storage-compute separation and Serverless, multi-mode compatibility, and intelligent security — together forming the complete picture of the GBase 8c kernel. Note that "one skeleton, two foundations, three layers of intelligence" outlines the trunk, while the eight focus areas are the specific implementation directions branching from it. They are not one-to-one mappings, but rather trunk and branches. In each subsequent installment, we will take one module and go into parameter-level and configuration-level detail.
Proven Kernel Capabilities
Architectural design must ultimately withstand validation in real-world scenarios and authoritative testing.
In May 2026, GBase 8c V6 passed the security and reliability assessment jointly conducted by the China Information Technology Security Evaluation Center and the State Secrecy Science and Technology Evaluation Center, joining the ranks of domestic distributed transaction-processing databases to earn this certification.
In August 2026, GBase 8c passed the HyBench benchmark with a high score of H-SCORE 3810.36, setting a new record and taking first place. In the distributed transaction-processing database category, this result is nearly 5× the previous best on the leaderboard.
This performance gain stems from architectural changes in GBase 8c V6 at both the storage and execution layers, enabled by the AI-Native Hybrid Execution Engine and reinforcement learning–based workload awareness:
Storage layer: introduces row-column hybrid storage with dynamic conversion, flexibly switching between row-store and column-store based on query patterns (hot data stays in row-store, historical data converts to column-store) — balancing transaction processing efficiency and analytical query speed.
Execution layer: optimizes execution plans and parallelizes operators; key operators such as aggregation and sorting support multi-threaded parallelism, with execution paths adapting to workload type.
Global Transaction Manager (GTM) and multi-tenant isolation: reduce cross-node transaction overhead and suppress resource contention between mixed workloads.
Today, these capabilities are deployed at scale in financial core trading, telecommunications billing and settlement, government GIS, and large model vector retrieval scenarios.
Breaking down the kernel is about seeing the database's true operating logic beyond the clutter of product specifications. GBase 8c's kernel is not only a natively distributed, multi-model, multi-form technical architecture — it is also a foundational answer to the demands of the AI-native era for high availability, high performance, and autonomous control.